G. Nebbia2, E. R. Dadashzadeh1,3, C. Rieser3, S. Wu1 1University Of Pittsburgh,Department Of Biomedical Informatics,Pittsburgh, PA, USA 2University Of Pittsburgh,Intelligent Systems Program,Pittsburgh, PA, USA 3University Of Pittsburgh,Department Of Surgery,Pittsburgh, PA, USA
Introduction: Since 2002, the liver allocation policy for adults is based on the Model for End-stage Liver Disease (MELD). While MELD was not originally created for this purpose, given its ability to predict short-term mortality, it has been serving as an urgency-based mechanism for organ allocation. Aiming to improve the MELD criteria, the purpose of this study is to investigate a data-driven approach using machine learning (ML) techniques to build a predictor of mortality for patients awaiting liver transplantation.
Methods: We retrospectively used the Scientific Registry of Transplant Recipients (SRTR) dataset, which included patients waitlisted for liver transplantation from 1985 to 2017, and we divided the dataset into three survival cohorts (i.e., 3, 12, and 24 months) including 88,758, 63,205, and 53,361 patients, respectively. We applied three ML algorithms (Logistic Regression, Random Forests, and Neural Networks) to predict the survival for each cohort, by training each ML model using 30 clinical factors such as functional status, additional laboratory values, diagnosis, blood type, BMI, and MELD itself. We removed patients that have substantial missing data in these factors, resulting in the final three cohorts of 25,560, 17,295, and 14,203 patients, respectively. For each cohort, 75% data were used for training and the rest unseen 25% for testing, measuring the prediction performance by the Area Under the ROC Curve (AUC). We analyzed each cohort as a whole and also grouped patients based on their specific diagnosis categories for sub-group analysis. The diagnosis categories we analyzed are Acute Hepatic Necrosis, Cholestatic Liver Disease/Cirrhosis, Malignant Neoplasm, Metabolic Disease, Non-Cholestatic Cirrhosis, and Other. AUCs of different models are compared by Delong test to assess statistical significance.
Results: MELD alone reached an AUC of 0.87, 0.78, and 0.75 for the 3, 12, and 24-month cohort, respectively. Logistic Regression reached AUC of 0.89, 0.83, and 0.82, while the other two ML models performed comparably. All the AUC improvements over the MELD baseline were statistically significant (p<0.05). In sub-group analyses, the AUCs of diagnosis-specific models showed consistent improvement over the sub-groups; in particular, the largest increase in AUC is achieved on the 24-month cohort for the diagnosis of Malignant Neoplasm, Non Cholestatic Cirrhosis, and Metabolic Disease.
Conclusion: This study shows that data-driven ML modeling outperforms the MELD criteria in predicting mortality for patients awaiting liver transplantation. We see a larger improvement (0.82 vs 0.75) when predicting a longer survival (24 months) and a smaller improvement (0.89 vs 0.87) for the 3 months. More promisingly, the improvement over different sub-groups indicates the ML models may be particularly beneficial to certain group of patients with specific diagnosis, potentially enabling precision prediction of survival on stratified patient cohorts.